ll

Review
The role of artificial intelligence
in the mass adoption of electric vehicles
Moin Ahmed,1,3 Yun Zheng,1,3 Anna Amine,1,2 Hamed Fathiannasab,1 and Zhongwei Chen1,*

SUMMARY
The electrification of mass transportation is hailed as a solution for Context & scale
reducing global greenhouse-gas emissions and dependence on un- Current global energy production
sustainable energy sources. The annual sales of electric vehicles relies heavily on unsustainable
(EVs) has continued to rise since 2011, with a global sale of EVs fossil fuels (coal, oil, and gas), and
of 2.1 million in 2019. This increase in sales is mainly due to the use of these fuels poses
continued improvement in the cost and performance of commercial serious problems: climate change,
EVs, increased EV options available to consumers, and environ- an increase in geopolitical
mental awareness. However, despite the positive outlook, EVs still tensions, depletion of these
face major limited resources, and adverse
challenges that hinder their rapid and widespread adoption: limited health effects. As a result, major
driving range, long charging times, and a lack of sufficient charging global environmental groups and
infrastructure. This review outlines the recent advances in EVs and governments have setup
related infrastructure, mainly from artificial intelligence (AI), which initiatives and plans to reduce
makes EVs a more attractive consumer option. The application of carbon emissions and combat
AI in improving EVs, facilitating EV charging stations, and EV inte- climate change, for example, the
gration with the smart grid is critically analyzed and reviewed. Paris Agreement. Renewable
Finally, future trends and prospects in the area are discussed. power generation and the
electrification of transportation
INTRODUCTION have seen massive research and
commercial developments over
Currently, global energy generation and transportation rely heavily on unsustainable
the last decade. Despite the
and nonrenewable fossil fuels, namely coal, oil, and gas. Apart from the limited sup-
numerous environmental and
ply of these fossil fuels, they also pose serious climate problems, geopolitical ten-
performance advantages of
sions, and health risks.1,2 In response to the detrimental effects of fossil fuels for
current commercially available
energy generation and transportation, governments have proposed initiatives,
electric vehicles, electric vehicles
such as the Paris Agreement, to curtail its use.3 The transition to renewable and
still have a small market share in
cleaner energy sources, such as solar, wind, and geothermal, for energy generation
the automotive industry.
and the electrification of transportation has been deemed an effective solution to
Compared with conventional
our reliance on fossil fuels and associated emission problems.4 The last decade
internal combustion vehicles, this
saw a rise in the sales of electric vehicles (EVs), with global sales of 2.1 million EVs
lack of consumer interest in
in 2019.5,6 There has also been a rise in electric automotive manufacturers, which
electric vehicles (EVs) lies mainly in
has given a wider range of EV options for consumers. Although Nissan and General
higher upfront costs, limited
Motors were among the few US manufacturers in 2012, major existing and new man-
driving range, and lack of
ufacturers, including BMW, Chevrolet, Tesla, Volkswagen, Renault, Honda, Skoda,
facilitating infrastructure.
and Opel have joined the race for EV sales in 2020. Growing consumer interest in
EVs can be attributed largely to the advent of new charging infrastructure, aware-
Regarding high upfront costs, it is
ness of EV environmental benefits, increase in EV performance and design specifica-
essential to bring the battery
tions, and the cost reduction of EV battery packs.7,8 However, despite the positive
pack’s cost down in the EV.
outlook of EV integration in society, EVs still face challenges to their widespread
Machine learning provides a cost-
adoption. These challenges include EVs’ limited range and the associated user
and time-effective approach to
range anxiety, lack of charging infrastructure, the high upfront cost of EVs compared
discover low-cost and high-
with that of traditional internal combustion engine vehicles, and safety concerns.9,10
performing battery materials and
Therefore, developing related novel techniques and proposing useful strategies are
necessary to overcome those challenges.

2296 Joule 5, 2296–2322, September 15, 2021 ª 2021 Published by Elsevier Inc.
 ll
Review

Recently, artificial intelligence (AI), which is defined as algorithms supporting models

aimed at mimicking natural thinking, perception, and action,11 has seen industrial increase battery-manufacturing
and academic applications in the field of EVs and related infrastructures, such as in efficiency. Artificial intelligence
EV battery design and management, charging stations, and even the smart grid.11–14 (AI) algorithms and controllers can
Within this review’s context, AI algorithms encompassing fields of machine learning provide a realistic driving-range
and computational intelligence are considered. Depending on the problem, these AI estimation and optimize energy
algorithms can outperform classical rule-based systems (also called expert systems), conservation, which can add extra
which uses human knowledge to define rules within a system, due to their ability to driving range and reduce
find unusual trends and simpler implementation.15,16 Several attractive advantages of consumer ‘‘range anxiety.’’ Access
AI application include: (1) EV cost reduction through optimal battery-material design to charging stations and the smart
and manufacturing,17,18 (2) accurate range estimation to mitigate EV consumer range grid can significantly improve the
anxiety by predicting future driving conditions,19 (3) improved EV energy consumption consumer appeal of electric
versus traditional controls using AI controls for the EV auxiliary systems,20 (4) a potential vehicles. Vehicle-to-grid (V2G)
for increased road safety and optimal traffic flowthrough connected and autonomous technology in the smart grid
driving,13,21 and (5) a thorough and efficient modeling approach for optimal location provides EV owners with an
and resource allocation for electric vehicle charging stations (EVCSs) and energy sched- opportunity to earn revenue by
uling for EV interaction with the smart grid.22 providing power to the grid. AI
algorithms make it easier to
For a better understanding of this topic, the AI techniques mainly used in EVs and manage EVs and power-
related infrastructures are summarized and divided into machine learning (ML) and generation interactions by
computational intelligence (CI), as shown in Figure 1. To be specific, ML models estimating optimal locations and
are adept at finding relations and trends between inputs and outputs based on pre- resources for charging stations,
vious observations, and they require training on a previous dataset. The ML used in reducing congestion in charging
EVs and related infrastructures can be roughly classified into supervised, unsuper- stations, and managing energy
vised, and reinforcement learning (RL). The supervised and unsupervised learning distribution in V2G technology.
have been researched and applied in areas within EV and its infrastructure where
large datasets are available or can be created, such as in EV battery-state estimation
and discovery of new materials for the EV batteries.23,24 Their basic concepts are dis-
played in Figures 1A–1C, in which the deep-learning (DL) model uses architectures
involving neural networks (NN) with more than one hidden layer25; RL aims to learn
the best course of action on its own through trial and error, and the agent interacts
with its environment through actions and is rewarded accordingly.26

Aside from ML, CI algorithms (as shown in Figures 1D and 1E, are commonly used for
solving search, optimization, and other complex problems.27 Within the EV context, CI
algorithms are instrumental in solving complex, dynamic optimization problems, such
as the optimization of control systems within EV, optimal placement of EVCSs, and inte-
gration of EV infrastructure with the smart grid.28–30 Although the current EV market has
not yet fully embraced AI, the number of related academic publications, patent applica-
tions, and the production scale have increased rapidly in the past decade. An overview
specifically for the roles of AI in EVs and their infrastructure is rare but is important to
accelerate the progress of R&D and mass industrial application and commercialization.

This review focuses on those areas within EV and its infrastructure where AI can
impact EV’s mass adoption. The outline of this review is illustrated in Figure 2; aside
from the basic introduction above, the use of ML and CI in battery manufacturing and
1Department of Chemical Engineering, University
management within EV and EV auxiliary control systems are discussed first. Further-
of Waterloo, 200 University Avenue West,
more, the role of AI in the optimal location selection and resource allocation of an Waterloo, ON N2L3G1, Canada
EVCS, and seamless integration of EV and the smart grid are considered in the sub- 2Hinsdale Central High School, 5500 South Grant
sequent two sections. Finally, future research orientations are also provided to Street, Hinsdale, IL 60521, USA
address challenges such as high-throughput data generation, performing in situ cal- 3These authors contributed equally
culations, and the scalability of data from battery-level to battery-pack scale, for *Correspondence: [email protected]
further research, technology applications, and commercialization. https://doi.org/10.1016/j.joule.2021.07.012

Joule 5, 2296–2322, September 15, 2021 2297

 ll
Review

Figure 1. AI techniques (ML and CI) used in EV and mass adoption

(A) Process flow for reinforcement learning. The agent interacts with the environment and gains a reward accordingly based on its action. Through
iterative learning, the agent learns and adopts a policy that aims to maximize its reward.
(B) Commonly used architecture for convolution neural networks (CNN). CNN, commonly used for image processing, as successive layers of
convolutional and pooling layers are connected to a fully connected neural network. Convolutional layers filter the image based on high-level image
features while the pooling layer compresses the image to reduce its size for ease of handling.
(C) Neural network of recurrent neural networks (RNN). RNN, commonly used for time series analysis, has the output of the neural network at (t-1) time-
step as the input of the at t time-step. The image shows a single RNN unrolled in time.
(D) Depiction of the process flow of particle swarm intelligence (PSO). PSO, commonly used for search and optimization engineering problems, has a
swarm of particles, which search in the solution space. The particles communicate both a particle’s personal best and the global best to find the
solution.
(E) Depiction of the process flow of genetic algorithm (GA). GA, commonly used for search and optimization engineering problems, has an initial
population of solutions. During each iterative process, the population undergoes mutation and crossover operations to find the solution.

AI IN ELECTRIC VEHICLES
The safety, reliability, and economic feasibility of EVs are critical for their mass adop-
tion, and those properties can be significantly improved through the implementa-
tion of AI. Applications of AI in EVs have been widely investigated and can be
roughly divided into several categories, including EV battery design, manufacture,

2298 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Figure 2. Overview of the use of AI in EV, EVCS, and EV integration with smart grid
This review covers the some of the application areas of AI for EVs, EVCSs, and the integration of EV with smart-grid systems. Within EV, AI has been
utilized for battery R&D to improve battery performance, battery-pack management, and energy management. For EVCSs, the use of AI for optimal
EVCS placement, congestion control, and reliable energy scheduling are discussed. Finally, the efficient energy management during the two-way
energy transfer between EV and the smart grid, enabled by grid-to-vehicle (G2V) and vehicle-to-grid (V2G), using AI techniques, are reviewed.

and management; range optimization; as well as EV control-system design and

optimization.24,31 In this article, the energy-related content, i.e., the use of AI tech-
nologies in battery management and range optimization for energy efficiency
enhancement, will also be a focus.

ML in battery R&D
As core components, batteries, especially the lithium batteries, play an important
role in providing the power source and energy storage for EVs. In an EV battery
pack, individual batteries are connected and assembled into battery modules, which
in turn are connected and assembled into a battery pack. The combination of series
and parallel connections of the batteries within a module and the modules within a
battery pack provide the desired potential and capacity.32 However, current EV bat-
teries still face performance-related issues resulting from barriers in battery design
and manufacturing to the battery management and optimization during operation
in EVs. Limitations in battery design and manufacturing lead to lower energy-density
of the EV battery pack, resulting in increased cost.33 Additionally, more energy-effi-
cient EV batteries can drastically alleviate user range anxiety. To achieve higher

Joule 5, 2296–2322, September 15, 2021 2299

 ll
Review

2300 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Figure 3. ML models for battery-material discovery and manufacturing-efficiency improvement

(A) Heatmap illustrating the available chemical heatmap and suitable potential materials discovered by ML techniques. Reprinted with permission from
Hautier et al. 46 Copyright 2010 American Chemical Society.
(B) Discrete Fourier transformation (DFT) calculations for energy storage and conversion. These calculations can be used to find relevant electronic
properties of the materials for ML training and testing dataset training. Reprinted with permission from Wu et al.47 Copyright 2019 Elsevier BV.
(C) An example of a battery-manufacturing process including slurry preparation and coating the slurry onto the current collector. Furthermore, the
examples of process controls of these subprocesses are shown that can be optimized and affects the intermediate product’s (battery electrode in this
case) characteristics (mass loading in this case).
(D) The ML-based regression model evaluation, which shows high accuracy in predicting the electrode mass loading (intermediate battery
characteristic) from the manufacturing process controls Reprinted with permission from Liu et al. 48 Copyright 2021 Elsevier.
(E) The interdependence of the intermediate product characteristics with the manufacturing process control from studying the ML models. Reprinted
with permission from Duquesnoy et al. 49 Copyright 2020 Elsevier.

energy efficiency, consumer perception, safety, and economic feasibility, ML tech-

niques to overcome the above challenges have received increased academic and
commercial attention. By facilitating battery-material discovery, characterizations,
and improving battery-manufacturing efficiency, ML can lead the way to higher bat-
tery energy efficiency and safety and improve consumer perception of EV range.

Coupled with advances in quantum mechanics (QM) and computational advance-

ments and the availability of materials datasets, ML provides a less capital- and
time-intensive approach for battery material discovery and characterizations than
traditional approaches.17,31,34 ML techniques effectively explore the chemical space
for new material discovery (Figure 3A) of higher energy density and safer battery elec-
trodes,35–38 solid electrolytes,39,40 and electrolyte addictives.41 Also, ML techniques
can account for complex molecular interactions, reaction pathways, and minimization
of side reactions.42 The dataset for the chemical and quantum mechanical properties
of known chemical species is crucial for new material discovery. Examples of publicly
available and comprehensive datasets include Materials Project,43 the Chemical
Space P roject,44 and QM9.42 ML techniques are frequently employed to predict
the anticipated properties of new materials from other known properties by exploit-
ing the trends of the existing known materials in the database. QM calculations or da-
tabases (e.g., QM9) are used to estimate the chemical properties, which are hard to
measure experimentally or unknown, of existing materials and these chemical prop-
erties are used as inputs to the ML techniques (Figure 3B). For example, QM calcula-
tions can estimate the redox potential and molecular band-gap information from
molecular composition. A suitable ML technique can be trained by using this informa-
tion along with other known relevant chemical properties, which can be used to esti-
mate the redox potential of currently nonexistent molecules.45

To further facilitate material discovery, ML can also automate, thereby removing hu-
man biases in electrochemical and material characterization techniques, such as elec-
trochemical impedance spectroscopy (EIS),50–52 cycle-life testing,50,51,53,54 and to-
mography.55 Although numerous research work has been demonstrated, ML-based
approaches to battery design have lacked commercial interest because of numerous
reasons, namely, insufficient dataset size,24 and the reluctance of the mature Li-ion
battery industry to change battery materials.56,57 Materials datasets used in research
work can be quite small, especially when hard-to-quantify material properties are
used.24 Small datasets affect data quality negatively and limit the ML’s model selec-
tion and accuracy. Moreover, shortages of government funding in material science,
long time frames for commercially viable material developments, and a lack of rele-
vantly skilled employees induces industrial hesitance to adopt new materials.57,58

Another strategy to improve battery energy density is to tweak the characteristics of

the intermediate product (e.g., battery electrodes) during battery manufacturing

Joule 5, 2296–2322, September 15, 2021 2301

 ll
Review

(Figures 3C and 3D), ensuring the battery’s energy density is close to its maximal
theoretical limits. For example, increased electrode surface area can lead to a higher
discharge capacity of a battery.59 Recently, some research works have used ML to
predict the intermediate physical product characteristics from manufacturing pro-
cess parameters.18,37,48,60–62 These process parameters can be chemical composi-
tions, such as during electrode active area fabrication or manufacturing equipment
parameters (e.g., the spinning speed of mixers). For instance, Lui et al. utilized an ML
regression-based model to predict the electrode mass loading from control param-
eters involved in slurry mixing (mass content, solid-to-liquid ratio, and slurry viscos-
ity) and coating onto the substrate (comma gap) during electrode fabrication.48 This
research area is relatively recent and suffers from the lack of readily available relevant
datasets. Researchers have to design and generate their datasets experimen-
tally,18,48,63 which is a time-, capital-, and labor-intensive process. The dataset
from the work of Liu et al. is publicly available,64 but has limited research usefulness
because of a lack of input features. More datasets might be available in the future as
the research and commercial work in this field increases. Simulations using physical
and empirical models can be used for quick dataset generation.60,49,65 For instance,
Takagishi et al. trained an artificial neural network (ANN) by using simulation results
from zero-dimensional electrochemical models, and this ANN can predict the
charge-discharge specific resistance from the electrode porosity features (porosity,
active material particle size and volume fraction, and compaction process pressure);
electrolyte conductivity; and binder/additive volume fractions.60 Although this
approach can generate a large dataset in a cost- and labor-intensive fashion, it
does not account for real-world phenomena, such as complex interdependencies
between manufacturing processes, and depends on the reliability of simulation re-
sults.49 Combining insights from experimental results into model-based simulations
can ensure high simulation reliability while generating large datasets for ML algo-
rithm training.49 Moreover, more studies need to study the interdependence be-
tween the manufacturing process and the output using insights from ML (Figure 3E).

ML in battery management
In EVs, the battery-management system (BMS) is responsible for battery-pack
sensing, battery-state estimation, and diagnosis and ensures energy-efficient control
of the EV battery pack (Figure 4A).66 BMS typically uses the voltage, current, and
temperature of each battery module to compute state-of-charge (SOC) and state-
of-health (SOH) for battery-state estimation and diagnosis, respectively.67 Battery-
pack SOC and SOH estimation are challenged by nonlinear battery characteristics
within each cell and inconsistencies in performance between them.14,67,68 The tradi-
tional SOC estimation method includes referencing from look-up tables, ampere-
hour integral methods, and model-based estimation methods, including equivalent
circuit models (ECMs) and electrochemical model (EMs). Similarly, SOH in EV has
traditionally been estimated by using ECM or other empirical models. However,
these methods for state estimation cannot predict the states accurately and/or
computationally expensive. Whereas EM-based models typically provide higher
SOC and SOH accuracy than ECM, EM calculations require high computational re-
sources, which makes them unsuitable for real-time EV applications.67 Here, the
term ‘‘computational resource’’ refers to the resource used by some computational
models in the solution of computational problems, such as memory space, amount
of storage, computation time, etc. Accurate health and power estimation are critical
for reliable and safe operation to prevent EV malfunctions and potentially serious ac-
cidents.12,14,69 ML-based state estimation is considered a promising approach for
EVs because of its lower computational demand, accuracy, and lack of need for
extensive mathematical models (Figure 4B).14,67,69 Moreover, a hybrid approach,

2302 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Figure 4. Machine learning (ML) in battery management

(A) The battery-pack sensor input information to the battery-management system. The battery-
management system collects the time series data, consisting of pack voltage, pack current,
voltages of individual cells, and temperatures from the temperature probes at intervals.
(B) Accuracy versus computational resources required for SOC and SOH estimation. When
measuring SOC and SOH, data-driven models, powered by deep learning algorithms, show higher
accuracy than the ECM models and much lower computational resources compared with physical
models, such as single-particle models.
(C) Schematic diagram of recurrent neural networks (RNNs).
(D) The temperature profile of the battery cells in a battery pack. Adapted from Karlsen et al. 71
under the Creative Common License [https://creativecommons.org/licenses/by/4.0/].
(E) The temperature profile of the batteries can be visualized as a histogram. 71 The temperature
variation can be trained by using an RNN to predict the future temperature values, which can be
compared with actual values for anomalies. 72 Adapted from Karlsen et al.71 under the Creative
Common License [https://creativecommons.org/licenses/by/4.0/].

in which ML and ECM/EM models are combined, can be used with higher SOC and
SOH prediction accuracy and low computational resource requirements.70 An over-
view of ML-based estimation methods for SOC and SOH is presented below.

As noted above, pretrained ML models can model the nonlinear dependencies to

state estimation without having high computational requirements as required by

Joule 5, 2296–2322, September 15, 2021 2303

 ll
Review

Table 1. ML approaches at different cell levels

Level Sub-Level Aim ML Input ML Output
Cell cell material design improve battery chemical property electrochemical property
energy density, from experiments or
safety, and cost DFT calculations
(<100 USD/kWh)
cell manufacturing battery cost process parameter process quality
reduction
(<100 USD /kWh)

System battery-management SOC individual battery SOC

system SOH terminal voltage SOH
temperature
fault detection faulty battery,
location, and time

DFT, density functional theory; SOC, state of charge; SOH, state of health.

EM. In the case of pretrained ML models, such as NN, the result usually requires sim-
ple matrix multiplication (the matrix containing weights and biases in the case of
NN), whereas EM models require simultaneous numerical solution of nonlinear par-
tial differential equations.25,73 Publicly available battery dataset sources useful for
SOC/SOH estimations include the NASA Acme dataset,74 CALCE battery group,74
and Severson et al..53 Among other supervised ML algorithms, NN75 and SVM76
have been often used for SOC estimations. However, these algorithms have shown
to have low accuracy rates (typically around 2%). For instance, a two-layered NN with
a filter to remove noise shows an accuracy of just about 2%.75 However, employing
DL algorithms can significantly improve the SOC estimation accuracy.76 This accu-
racy in SOC estimation comes from increasing the number of hidden layers, but
overfitting the data can become a challenge. In particular, recurrent neural networks
(RNN)-based models can be useful as it can consider the battery history and evaluate
the battery’s dynamic aging and hysteresis (Figure 4C).77 Similarly, RNN has also
been successfully employed to estimate SOH.78 Types of RNN architecture,
including RNN-gated recurrent units (RNN-GRU) and long short-term memory
(LSTM), can effectively capture the long-term battery characteristics. The inputs
for ML algorithms in SOC and SOH estimations are listed in Table 1. Hybrid models
have shown SOC prediction, with an error of less than 1% in some cases,70,79 while
requiring much less computation resources than ECM/EM models.69,70,79 However,
for real-time EV applications, the computational power of the current BMS in EV
needs to be considered. Further challenges to ML adoption for real-time SOC and
SOH estimation are: (1) the inability of pure ML models to model battery aging
mechanisms, (2) a wide variety of external EV operating conditions that affect battery
degradation, and (3) the fact that most research work is conducted at the cell level
rather than at the battery-pack level.64

Safety is a critical topic in battery applications and another operation of BMS is to

minimize the risks of battery failure during EV operations, which can be life threat-
ening in the case of battery-pack runaway.80 Early battery-fault detection using ML
approaches has seen recent development in EV battery-pack diagnosis.81–83 Early
detection of location, time, and cause of a battery cell or module failure in an EV

2304 Joule 5, 2296–2322, September 15, 2021

 ll
Review

battery pack ensures user safety and EV longevity. Faults in battery cells can be
caused by operational abuse (overcharge, overdischarge, and extreme temperature
exposure), faulty external connections, and mechanical damage.84,85 Although
operational abuse can be overcome by having electrical charging controls and a
thermal management system, EV battery packs are still prone to mechanical defects
from operations, such as vibrations during trips. Traditionally, in BMS applications,
the faulty battery location is determined by comparing the current battery-terminal
voltage with its historical estimators (such as mean and standard deviation)86,87 or
voltage across the load88,89 (Table 1). However, these approaches are not suitable
for real-time EV applications mainly because the EV voltage profile during battery-
pack charge and discharge is dynamic and nonuniform. It is interesting to note
that the above-mentioned mechanical faults can be detected by changes in bat-
tery-terminal voltage and surface temperature.85,90 DL algorithms have been suc-
cessfully applied for voltage or temperature changes and for faulty battery time
and location by using relatively low computational resources.81,72

In case of terminal voltages, statistical measures for variations between cell voltages
(such as voltage differences and covariances) and within cell’s historical voltages
(such as rolling variances) can be used as inputs for DL algorithms. Specifically, the
general regression neural network (GRNN) has been employed for fault detection
with a high accuracy (>95%).81 Meanwhile, when battery surface temperature (Fig-
ure 4D) in the battery pack is used as a fault indicator, RNN can be trained to predict
the normal battery temperature. The prediction from this trained model can be used
to compare the actual values for anomalies (Figure 4E).72 Therefore, compared with
traditional methods, methods based on DL are more applicable to potential real-
time EV use because of its accuracy and ability to facilitate changing load profile dur-
ing EV use. However, insufficient battery-pack-level datasets with a wide variety of
operating and external conditions and incomplete battery-fault knowledge currently
hampers the deployment of these ML algorithms for commercial purposes.
Acquiring knowledge of battery faults requires experimentations with numerous
experimental conditions and samples. Again, ML can be used to find trends between
the cause and consequence of a battery fault by using datasets from experiments
and physical models.91–93

ML in range optimization
Besides the battery-management technologies reviewed above, range optimization
during EV driving can tremendously increase energy efficiency and save driving
time. In general, range estimation (RE) is a step to achieve EV range optimization.
As one of the most important research areas of EV today, an accurate RE can largely
mitigate range anxiety faced by the EV drivers because of limited driving range.94,95
EV drivers can make effective driving, parking, and charging decisions and partici-
pate in more vehicle-to-grid (V2G) charging. However, conventional RE methods
are not always accurate mainly from its disregard of dynamically changing external
and operational conditions.96,97 For example, the range estimators in Tesla’s Model
S use the energy consumption of the past few miles to estimate the future available
range, without considering changes in environmental, driving conditions, and
behaviors.94

Compared with traditional RE methods, AI can provide opportunities for accurate RE

by mapping the complex interaction between RE and the factors affecting it. Further-
more, AI algorithms, including ML, can be used to accurately predict the future envi-
ronmental and driving conditions based on past and current conditions. These future
predicted conditions can be used for a more accurate RE. Table 2 summarizes some

Joule 5, 2296–2322, September 15, 2021 2305

 ll
Review

Table 2. Representative research of machine learning (ML) in EV range estimation (RE)

Approach Parameters ML algorithm RE accuracy
Historical data battery significant parameter 1.63 km (MAE)106
SOC voltage (min, max) identification using
current (min, max) correlation analysis
temperature (min, max) followed by multiple
vehicle linear regression (MLR)
speed (avg) classification and regression 1.27 km (MAE)98
external tree (CART)
temperature
visibility gradient boosting decision 0.82 km (MAE)98
tree (GBDT)
precipitation
battery artificial neural 2.2% (MSE) accuracy for a
SOC networks (ANN) 50.4 km real-life EV trip103
SOH vehicle
auxiliary load
weight
external
road type
traffic
temperature
driving behavior
Predicting future energy/power Vehicle speed linear regression (LR) 2.18 km (MAE)105
consumption recent energy consumption support vector 1.95 km (MAE)105
external regression (SVR)
road elevation
vehicle MLR 1.95 km (MAE)104
speed principal component 2.07 km (MAE)104
acceleration regression (PCR)
past power consumption
clustering of data using self-organizing maps ( 0.70 km (MAE)104
past distance
SOM) followed by
past trip run time
regression tree
temperature
weight of loads
tire pressure
frontal area
external
road elevation
Abbreviations are as follows: MAE, mean absolute error; MSE, mean-squared error.

of the relevant works in the literature for RE using AI. AI algorithms have been used for
RE by directly using environmental and historical driving behavior data,98,99 esti-
mating the EV battery energy or power consumption,97,100–103 and identifying
driving conditions and behaviors.97,104,105 For ML training, real-time historical
discharge EV data are processed to extract relevant battery, and vehicle, and external
parameters (Table 2) and to remove missing and erroneous data. Moreover, the EV
historical data can be fused with historical weather, traffic, and road conditions to
include external parameters, thereby increasing the RE estimation accuracy through
ML training.98–100 Parameters that have most impact on RE, such as battery SOC and
external temperature, can be estimated by using correlations and are considered to
reduce ML model complexity and training time.98 ML models can be trained to
output EV range directly, as well as the EV future energy or power consumption. By
considering dynamically changing EV internal and external conditions in a computa-
tionally resourceful manner, and without use of complex explicit models, ML allows
for an opportunity for more accurate RE.

The EV control system

The EV control system also plays an essential role in EV hardware’s energy consump-
tion and thus can indirectly impact the EV’s driving range. Optimal EV control-system
use ensures reduction in energy consumption of EV hardware including power steer-
ing, regenerative braking, and internal environment hardware controls such as HVAC
(heating, ventilation, and air conditioning) while maximizing vehicle speed.107

2306 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Figure 5. EV control architecture and process flowcharts of some related controllers

(A) EV control architecture.
(B) Conventional (PID) controller.
(C) PID-artificial intelligent controls (AIC) (particle swarm optimization [PSO] based) controller.
(D) AIC (fuzzy-logic based) controller.

Particularly, the regenerative braking system (RBS), available in most EV models, pro-
vides an opportunity for energy generation for increased driving range. The RBS con-
verts kinetic energy into useful electric energy when the vehicle slows down, thereby
extending the range of the EV.108 Driving technique, system temperature, and
ambient temperature affect the efficiency of energy extracted from RBS.109 Figure 5A
shows the energy flow and communications between different EV components
including charging system, BMS, electric propulsion system, and auxiliary systems.

Joule 5, 2296–2322, September 15, 2021 2307

 ll
Review

Control systems involve controllers that act to change the state or process of the sys-
tem based on the inputs from the sensor readings. Artificially intelligent controls
(AICs), involving AI techniques such as fuzzy logic, NN, and evolutionary algorithms,
can be either used as a substitute for or in conjunction with conventional industrial
controllers, such as proportional, integral, and derivative (PID) controllers (Fig-
ure 5B).110,106 Among current optimization strategies for EV control systems, AIC,
either exclusively or as an AIC-PID hybrid model, are researched as an emerging
and smart choice for EV control-system design and optimization to improve the en-
ergy efficiency and further relieve range anxiety.28,108,109,111–116

As for hybrid AIC-PID controller, CI algorithms, including particle swarm optimiza-

tion (PSO) and ant colony optimization (ACO), can be used to tune the critical pa-
rameters of proportional (Kp), integral (Ki), and derivative (Kd) parameters to further
reduce the process steady-state error, overshooting.117–119 A case in point is the
PSO algorithm used in hybrid PID-AIC controller (Figure 5C), in which a predeter-
mined number of random solutions (i.e., PID parameters) are initialized in the
search space. Each solution is evaluated, compared, and updated with others to
search for the optimal one. Similarly, ACO attempts to find the optimal path in
the search space by having the solutions track their path. The optimal solution
leaves the most intense pheromones trail, which influences the path of other solu-
tion. As a result, PID parameters tuned with PSO and ACO algorithms have been
shown to aid in reducing the assisted current drawn by the electric power-assisted
steering (EPAS), further translating into better EV energy conservation.28 Further-
more, using NN to tune RBS PID controller has a positive effect on PID’s response
time, steady-state error tracking, and resisting perturbation, resulting in higher EV
driving range.115

Regarding the AIC controller, a typical one is the fuzzy-logic-based AIC controller,
which is also called a fuzzy-logic controller (FLC). As shown in Figure 5D, fuzzification
is the process to convert numerical input signals into linguistic equivalents, and de-
fuzzification uses rule-based inference, a linguistic output determined and trans-
formed into a numerical output.120 In the EV context, FLCs are deemed superior
to conventional controllers in HVAC systems because they can effectively handle
user comfort while reducing energy consumption.111,121 Moreover, this reduction
in energy consumption results in higher EV driving range when SOC and vehicular
speed are used as FLC inputs.111 Regarding the specific AIC applied in RBS, FLC
and NN have been researched with respect to the braking allocation between regen-
erative and conventional hydraulic braking systems. This braking allocation ensures
maximum energy generation while maintaining EV user safety. Braking (e.g., braking
pedal displacement), battery (e.g., SOC and temperature), and vehicular speed can
be used as FLC inputs to optimize the braking allocation for highest energy conser-
vation.112–114 When tested on a prototype EV, the FLC system showed an improve-
ment of 16 %, 22.2 %, and 25.7 % in motor efficiency, energy efficiency, and
maximum driving range, respectively, compared with that of the nonfuzzy RBS.114
However, for real-time EV applications, it is important that the hyperparameters of
NN are optimized for functionality and processing speed.122 Alternatively, the FLC
can trigger the RBS motor in case of driver coasting or creeping behavior, when
representative FLC inputs, such as the stokes of the acceleration and braking pedals,
are used. Results from the simulations and hardware-in-the-loop tests demonstrated
a higher energy regeneration and braking stability.116 Although in FLC rules are
clearly defined, the construction of these rules is nonobvious, difficult, and requires
expert knowledge. In such a case, AICs based on CI algorithms can be used to ease
the control design and performance.122

2308 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Figure 6. Cases for optimal placement of electric vehicle charging stations (EVCS)
(A) Tesla superchargers consisting of distribution panelboard, metering, and incoming power
source. 123
(B and C) Clustering of spatial map based on traffic density and EV driving distance. For traffic-
density distribution, the traffic data of road segments within a predetermined grid are aggregated
and overlaid on the corresponding grid. As shown, the traffic density is pictorially represented by
the radius of the red circle. For driving-distance clustering, the regions on the map are clustered
based on the location of EV and its destination points. 126
(D) Objectives of EVCS are modeled by using as a MOOP problem, which is solved by using CI
algorithms (PSO)
PSO was used to solve for MOOP, which considers land-cost and distribution investments;
meanwhile running cost was considered as the restraint. Solutions of PSO lead to Pareto solutions
of optimal EVCS placement. Reprinted with permission from Awasthi et al. 127 Copyright 2017
Pergamon.
(E) Agent-based modeling for EVCS optimal location. In this example, the agents (EV owners, EV
drivers, and EVCSs) are placed in a geographical environment and their attributes (mentioned in
the figure) are modeled. The two-sided arrows represent the interactions between the agents.

AI IN ELECTRIC VEHICLE CHARGING STATIONS (EVCS)

Prediction of the optimal location
One of the important ways to reduce the EV user range anxiety is the optimal
placement of EVCS, namely the optimization of charging locations. This placement
will give the user more confidence in using an EV, given that the user has a higher
chance of finding an EVCS before the EV battery dies. EVCS are classified based
on the charging voltage provided by the charger, and these categories are Level
1 (120 V), Level 2 (240 V), and DC fast charging (about 400 V in case of Tesla super-
charging stations) (Figure 6A).123,124 To gain wider EV acceptance and provide
user convenience, Tesla has installed about 1,971 DC fast charging stations glob-
ally. Level 1 EVCS, with long charging times, are usually placed in parking lots,
wheras DC fast EVCS are placed near high-volume road access points. Another
company is ELaadNL, which is one of the largest Dutch EV-infrastructure providers,
provides optimally placed charging spots throughout the Netherlands and

Joule 5, 2296–2322, September 15, 2021 2309

 ll
Review

promotes the use of data analytics, including the use of ML, to plan for EV
infrastructure.125

Optimal placement of EVCS depends on several factors, such as the local charging
demand, construction feasibility, road network and other infrastructure, operating
economy, and power-grid security.128 Optimal placement of EVCSs is generally
formulated as a multiple objective optimization (MOOP) function with objectives
comprising minimization of costs, maximization of net present value (NPV), or pref-
erence to unpopulated areas.126,127,129–133 The cost function, which is beneficial for
the EVCS owner and from the policymaker’s perspective, includes the costs of EVCS
construction, operations, and charging.127,134,135 The total costs of EVCSs are
formulated into annual costs by considering the expected total time of EVCS oper-
ations and market interest rates.134 NPV, an economic evaluator, discounts future
cash flows from a project (an EVCS project in this context) and, together with the
costs, indicates the economic feasibility of the EVCS project.136 Therefore, maximi-
zation of NPV and minimization of total costs of an EVCS project are crucial for
policymakers, investors, shareholders, and owners of the EVCS projects. Moreover,
Pevec et al. examined the objective function, which gives EVCS placement priority to
locations where there is a smaller number of existing charging stations.129 By popu-
lating EVCS in low EVCS densities, the convenience of EV charging for EV owners in
these areas is enhanced, and the local municipal government’s objective of EVCS in-
stallations are met. The primary purpose of optimal placement using MOOPs is to
identify optimal EVCS placement that satisfies the objective function.

ML is used for preparation of data or models for these MOOPs,129,126,137 and CI,
including swarm intelligence (e.g., PSO) and evolutionary algorithms (e.g., GA)
can be used for solving these MOOPs.127,130–133 At the data and model preparation
stage, supervised ML models, as opposed to mathematical models, can be trained
on a dataset containing data on existing real-time EV operation, traffic, and EVCS
operations, and hence, the predictive ability of this approach is more in tune with
real-world scenarios. As for the latter, in CI, the use of evolutionary computing algo-
rithms for solving charging-station placement problems has also been widely inves-
tigated.133 The solution approach for optimal placement of EVCS can be divided
into three main steps, which include data collection (Figures 6B and 6C), data pre-
processing, and placement improvement. ML clustering algorithms, such as hierar-
chical or K-means clustering, can be used to form zones, within a geographical
context, based on the typical EV driving distance and destination locations.129,126
In the second step, ML methods are used for modeling EV user demand and traffic
occupancy, which are subsequently used in MOOP.30,129 When it comes to the third
step, optimal placement is determined by solving for the MOOP by using CI algo-
rithms such asas PSO127 and GA130 (Figure 6D).

Aside from the MOOP relevant approach, ML can be used for optimal EVCS site
placement without formulating a MOOP. This approach removes the calculations
required in the MOOP case, which allows for the inclusion of qualitative data, re-
moves underlying models and corresponding assumptions, and possibly removes
the computational load required for MOOP.138 EVCS placement without MOOP
has been solved by using supervised ML algorithms such as K-means clustering,137
Bayesian networks,138 and NN.135 In these models, where ML models are exclu-
sively utilized, the ML models are trained on real-life datasets and output the
preferred EVCS location. Hence, despite the stated advantages of this approach
over MOOP, a large dataset is required to ensure model robustness and reliability.

2310 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Another research approach to determine EVCS placement and sizing is through

agent-based models (ABM). In this approach, different agents (such as EV owners,
EV drivers, and EVCSs) are assigned different attributes and interact with each other
in a model environment within a geographical location (Figure 6E).139–141 Despite
ABM’s complexity and higher computational demand compared with MOOP, it al-
lows for a more realistic modeling tool by considering different agents, their attri-
butes and their complex interactions.141,142 Furthermore, the agent’s attributes,
such as the charging behavior of the EV driver, can be modeled with stochastic pro-
cesses to account for different possible real-life scenarios. Within the context of
EVCS placement, ABM has been used to model and study the short-term and
long-term impact of EVCS placement and charging capacity on EV adoption and
charging use, which allows for EVCS roll-out strategies.139 ML techniques, both su-
pervised and unsupervised, can be implemented in ABM to improve model accuracy
by modeling agent behavior and decrease the computational demand.143,144 With
the advent of charging datasets (publicly available from ElaadNL145 and CAN-
Data146), the EV owner’s charging behavior can be modeled by using supervised
and unsupervised ML algorithms, including DL.147 Furthermore, the ABM’s input
(the agents’ inputs, attributes, and interactions) and output can be trained and sub-
sequently modeled by using supervised ML algorithms, which reduces the overall
computational load requirement and aids in the ABM model’s understanding by
replicating the nonlinear behavior.143,148

Energy scheduling and congestion management

Compared with building new infrastructure, which requires significant planning,
investment, and legal agreements, another energy- and time-efficient approach to
increase consumer interest in EVs is the efficient use of existing EV-charging infra-
structure. Accordingly, AI-based routing algorithms can be used to further improve
the efficient use of existing infrastructure such as EVCSs via energy scheduling and
congestion management. Therefore, this section deals with the AI application in
strategies that aim to improve resource- and usage-efficiency, while considering
user comfort and convenience. Specifically, energy forecasting and management,
congestion reduction, and fast and smart charging within an EVCS are mentioned.

For the EVCS to smoothly operate within the energy constraints of the local grid, it
needs to balance the EV-charging demand and available load supply. Hence, it is
important to predict the energy and charging demands of EVCSs and to use the en-
ergy resources efficiently. The steady penetration of EVs in the automotive industry
has led to the availability of user EV-charging data. ML approaches on this EV data
combined with other temporal (e.g., temperature) and geographical data (e.g.,
points of interest) can lead to a more realistic and accurate EVCS energy and
charging demand estimation. Furthermore, compared with the traditional determin-
istic approach to energy determination, the ML-based approach accounts for uncer-
tainties resulting in real-world charging applications by using relevant historical
datasets.149 Within this context, ABM can be used to identify EVCS capacity on
the basis of available energy resources, availability of EV drivers, and their charging
behaviors.139,150,151 ML algorithms, including linear regression152 and NN,153 can be
used to predict the charging behaviors of EV consumers and estimate the energy de-
mand.152–156 ML methodologies used in these works are general and applied to any
specific geographical location for accurate EVCS demand and charging predictions.

For EVCS user convenience and comfort, it is important to reduce the queue and
charging times. It is worth noting that, even with fast DC charging, it takes
30–60 min to charge an EV fully, whereas it takes minutes to fill the tank in a traditional

Joule 5, 2296–2322, September 15, 2021 2311

 ll
Review

Figure 7. Examples of EVCS congestion management using RL and smart charging using ML
(A and B) RL for congestion management. The map of three EVCS is mapped into a grid where each
EV starts from the same starting point and the relative distance of EV to EVCS is maintained.
Furthermore, the traffic density of the road segment is mapped onto the individual grid. The RL
algorithm optimizes the congestion policy to minimize the total EVCS waiting time. 157
(C) Bayesian-based ML model for fast-charging protocol
In this iterative ML-based charging protocol, the outcome of the battery is determined by using a
predetermined ML algorithm. Based on this early outcome prediction, the Bayesian optimization
determines the next charging outcomes for successive cycles in the four-stage MCC charging
protocol. The Bayesian-based ML algorithm, along with the ML-based early-life predictor, finds the
charging protocol, i.e., the C-rates, for the first three cycles.54

internal combustion engine vehicle. Therefore, an additional EV in a fully occupied

charging station can add the waiting time for other cars in the queue for about half
an hour. The use of supervised and unsupervised learning in congestion manage-
ment in EVCS placement is limited because of a lack of limited relevant EVCS data-
sets. As a potential solution, RL can be used for appropriate policy determination
to schedule EV to relevant EVCS to minimize wait times (Figures 7A and 7B).157

Apart from congestion management, AI algorithms are used for the determination of
optimal charging protocols without compromising safety and EV battery-pack life in
academic and commercial research.158–160 The common charging modes include
constant current–constant voltage (CC-CV) and multi-stage constant current
(MCC). In CC-CV, the batteries are charged to cut-off voltage at a constant current
and then maintained at cut-voltage. Meanwhile, batteries are charged in intervals in
MCC, where different currents are supplied at different time intervals. The current is
held constant within a time interval. In the research community, MCC is a popular
choice for fast charging, and the research efforts are focused on determining the
best current and time intervals for charging.161 It would be extremely time
consuming to test for different numbers of charging time intervals and the combina-
tions of charging current for these time intervals by using physical experiments. As
an alternative approach, the charging current for each time interval can be formu-
lated as using a model, which can be solved by using evolutionary algorithms.162–166

2312 Joule 5, 2296–2322, September 15, 2021

 ll
Review

More recently, Attia et al. used ML for current determination for the fast charging of
lithium-ion batteries while setting the number and time of charging stages. An early
detection model, that can predict the lifetime of a lithium-ion battery from initial
cycling data, was followed by a Bayesian optimization algorithm to probe into the
possible search space (Figure 7C). As opposed to searching for C-rates for each
charging step, Bayesian optimization was used to find promising C-rates, of the first
three cycles of the MCC, based on results from the iterative process. The C-rate of
the fourth cycle is dependent on the first three cycles. As a result of this approach,
significant time- (560 versus 16 days) and cost savings were achieved when
compared with brute-force search because this ML-based approach optimally
searches the promising C-rates from the entire search space.53,54

AI IN THE INTEGRATION OF EV WITH THE SMART GRID

Introduction and main challenges of the integration
Smart grids, in comparison with the traditional electric power distribution systems,
allow for two-way energy flow, secure dynamic optimization of energy flow opera-
tions—such as determining the pricing of charging an EV based on the supply and de-
mand of electricity—and smoother integration of renewable-energy production and
storage.167–174 The direction of energy flow from the grid to EVs is referred to as
grid-to-vehicle (G2V), in the case of EV charging, and vehicle-to-grid (V2G) in opposite
energy flow.175 Therefore, from the smart-grid perspective, the EV can be viewed as en-
ergy storage during charging in G2V and potentially an energy generator during its
discharge to the grid in V2G. As shown in Figures 8A–8D, this bidirectional energy
flow allows for an efficient grid-energy generation and distribution through frequency
regulation, peak shaving and load leveling (Figure 8B), load regulation (Figure 8C),
and spinning reserve (Figure 8D), particularly in the case of grids wholly or partially pow-
ered by intermittent sustainable energy sources such as solar and wind.175–177 Further-
more, this attractive feature of energy generation to the grid can be a revenue stream for
an EV owner, and V2G technology has also garnered major industrial interest.178

On the other hand, regardless of the advantages and benefits, V2G and G2V also
face several technical, economic, legal, and social challenges, which include social
resistance to V2G, energy distribution complications, hardware barriers, and high in-
vestment cost.175,179 One typical challenge is the scheduling and distribution of the
smart grid with EV, which is complex because it requires considerations of several,
and often conflicting, objectives and restraints, including lower operational costs,
maximizing profits for power-generation plants and EV owners, minimizing carbon
emissions, and matching the real load curve with the target load curve. To deal
with those challenges, AI can be used as an effective tool. Specifically, AI algorithms
can regulate the energy scheduling and optimization problems resulting from the
complex two-way interaction between the EV aggregates and renewable-energy
generation systems. Moreover, AI can be also instrumental in ensuring smooth po-
wer distribution considering renewable-energy generations’ intermittency.

Optimization of power generation and distribution

Power generation and distribution are restrained by meeting the load demand and
supply, power-generation limits, voltage bounds, and power line thermal capacity.
To optimize power generation and distribution by consideration of those restraints,
both CI and ML have been investigated and applied as an effective strategy. As for
CI, the restraints can be first formulated as MOOP,29,175,180–183 and further solved by
evolutionary computing algorithms, which can handle multiple nonlinear MOOPs
with multiple restraints using reasonable computational hardware and software.175
For instance, Soares et al. optimized V2G and G2V scheduling by formulating a

Joule 5, 2296–2322, September 15, 2021 2313

 ll
Review

Figure 8. Integration of V2G with the grid and its effects on peak shaving and load leveling
(A) Bidirectional energy flow between smart grid and EV. Grid-to-vehicle (G2V) and vehicle-to-grid
(V2G) technologies allow for bidirectional flow of power between the EV and the grid. EV cab
utilizes the power from the grid when charging its batteries. Moreover, when EV is not being used
and has excess power, it can transfer that power to the grid. Moreover, this bidirectional power flow
can be integrated with the renewable-power generation systems to ensure sufficient power is
available in the grid.
(B) Peak shaving and load leveling. Batteries in the EV can be used as energy storage to ensure
target load is achieved in the grid. EV batteries can be charged at times of high power generation
which can be used at a later time to supply power to the grid at times of low power generation.
(C) Load regulation. The load profile can be regulated up and down to ensure that the same loading
is achieved. The EV can be considered as a dynamic load when it is charging. When the fixed load is
increased, Ev charging can be reduced (by increasing the power cost, for example) as in the case of
regulation up. In the opposite case of reduced fixed load, EV charging can be encouraged.
(D) Spinning reserve. In case of power outrage, addition power (spinning reserve) can be employed
to compensate. The dynamic load, which includes EV charging, can be reduced, which in turn
reduces the overall load.

MOOP to minimize total operation cost while maximizing the revenue obtained by
EV users. The demand and supply load matching, power generation, voltage and
thermal limits, and EV-charging limits were considered constraints. A Pareto front
of the MOOP in a simulation case study with a 33-bus distribution network and
1,800 EVs was solved by using PSO with good convergence.29 Additionally, Su
et al. solved an operational cost minimization objective function with the restraints
considering reducing EV battery life and EVCS queue waiting time. The former re-
straint was related to a battery-cost model that considers the battery SOC, and
the MOOP was solved by using an artificial fish swarm algorithm (AFSA). The model’s
optimized parameters were able to reduce the cost of battery degradation by 80%
and queue wait time by 60%.182

2314 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Aside from CI, ML can also be used in optimization of power generation, scheduling,
and distribution. A case in point was investigated by Yang et al., who formulated an
MOOP to minimize system power fluctuation and battery degradation, in which the
battery lifetime model is based on a DL algorithm, specifically LSTM, a type of DL
algorithm. Simulation results based on microgrid data in Belgium demonstrated
the algorithm’s ability to reduce both system power fluctuation and required EV
charge-discharge cycles for V2G.183 Moreover, Shipman et al. used ML-based
mode by using factors affecting users’ decision to participate in V2G and their reli-
ability in following through with their decisions. EV user availability and location
were tracked by using an app and formed the model inputs. The model was able
to predict an EV fleet’s availability at a charging spot for each hour during 4 weeks
with an error rate of only 5%. This prediction model is paramount in estimating
the available storage capacity.184

Optimization of renewable energy relevant systems

Similar energy management and optimization approaches can be applied to sys-
tems with renewable-energy generation systems, such as solar and wind power
plants. The electrochemical storage capability of EVs can in particular be used for
load shaving and load leveling due to intermittent power generation from renew-
able-energy sources.175–177 Here, example research works on the integration of
V2G with the power-generation grid with renewable-energy sources are presented.
Soares et al. used PSO to solve for a MOOP to minimize operation cost and maximize
EV owners’ profits. In a case study with a 32-bus distribution network and 50 plug-in
hybrids, PSO demonstrated to be 148 times faster than mixed integer programming
linear programming.185 Additionally, Liu et al. solved a MOOP by using PSO to mini-
mize the power grid operators and EV users’ cost, global CO2 emissions, and wind
curtailment when coordinating EV charging and discharging activities with the po-
wer grid of thermal plants and wind farms. The best results from the Pareto front
were determined by using decision-based fuzzy rules. Their results showed that
the EV discharging coordination at peak power loads with low winds could curtail
CO2 emissions.

Furthermore, EVs charging during times of low demand and high winds can reduce
the wind curtailments.180 To minimize the high variations in wind-energy generation,
Ghofrani et al. used GA and Monte Carlo simulations to coordinate the charging and
discharging behavior of EV fleets, based on their daily driving habits. Based on the
EV hourly data regarding parking time and distance, the EV fleets are grouped into 6
categories. In conjunction with Monte Carlo randomness in search procedure, GA is
used to solve for the MOOP on costs of wind imbalances and V2G expenses. The
simulation results on an EV fleet of 484 and a 10-MW wind-power station show a sig-
nificant merit in using V2G compared with gas-powered generators and battery sys-
tems to offset the wind-power imbalance.186

CONCLUSIONS
Key summary
Mass EV adoption is considered a necessary step toward decreasing global reliance
on unsustainable energy sources and reducing emissions detrimental to the environ-
ment. Recent years have seen wide commercial and academic research efforts to
overcome the challenges to EV adoption and make EVs attractive to consumers.
This review focuses on those efforts where AI has been applied, including the related
areas of EVs, EVCSs, and the interaction of EVs with the smart grid, in which commer-
cial interest in these areas have been summarized and even highlighted. Specifically,
AI algorithms used in EV battery design and discovery, battery management, and

Joule 5, 2296–2322, September 15, 2021 2315

 ll
Review

accurate RE, and the role of AI in the smart control of EV hardware and auxiliary sys-
tems to conserve battery energy. In the subsequent section, AI in optimal location
and energy usage of EVCS is discussed with consideration to EV user charging con-
venience and comfort. Finally, the AI role in the complex energy management and
optimization problems in V2G applications are reviewed.

Outlook
Overall, current research and practical commercialization of AI in EV and related
infrastructure are still in the early stages of development. Some areas of EV, where
AI is used are already being commercialized more than others. It seems that the
technical maturity of those techniques is sufficient to meet the requirements for prac-
tical implementation. To overcome current challenges, such as high-throughput
data generation, performing in situ calculations, and the scalability of data from bat-
tery-level to battery-pack scale, and further facilitate the development of AI tech-
niques used in EV and related infrastructure, some future orientations are provided
below.

AI to further enhance the processing capacities of battery management,

battery disposal, and energy controls
The processing capacity of the BMS is one of the limitations of using AI for real-time
battery-state estimation (SOC) and diagnosis (SOH and early fault detection). In prac-
tice, BMS design uses embedded microprocessors for computations.86 However,
with advancement in embedded systems from Moore’s law and scalability of ML al-
gorithms,25 ML-based approaches would be expected to be used for practical real-
time EV applications. Additionally, BMS computations relating to SOC, SOH, and
fault detection can be computed remotely on a cloud computational server, which al-
leviates the computational burden of BMS microprocessors.93,187 Moreover, ML
model training on battery-state estimation and fault detection have been conducted
on small and specific datasets by research communities. This training on small data-
sets results in model that tends to retain that specific dataset instead of the underly-
ing relationships. Implementation of such insufficient ML models on commercial
levels can have dire safety consequences, such as failure to detect potential thermal
runaway. As EV battery-operational datasets become more available with increased
EV use, there is an opportunity to research and implement battery-pack-level ML
models for real-time EV applications. To understand the underlying mechanisms,
identification, and detection of faults in batteries, iterative experiments followed
by ML for parameter exploration and exploitation can be conducted.93 Furthermore,
EV manufacturers can include ML-based fault detection algorithms that continuously
learn from historical EV data periodically, which can effectively consider specific EV
users’ driving habits, common driving conditions and external conditions (e.g.,
weather conditions).

Another potential risk of using AI in EV battery-management and control systems is

related to user privacy violations.188 As more AI algorithms are used in EVs, user
data—for example, driving habits and frequently visited locations—would be
collected and stored by EV manufacturers or other third-party vendors. Potential
EV users might be hesitant to share such information and this hesitance can lower
the overall EV appeal to the population. Also, this mass data sharing and storage
adds pressure for EV manufacturers.

Interestingly, another potential of AI approaches for remaining-useful-life (RUL) esti-

mation can be used for battery waste and disposal. Higher sales and consumer inter-
est in EVs raises a concern for the battery waste.189 A report from the National

2316 Joule 5, 2296–2322, September 15, 2021

 ll
Review

Laboratory of Renewable Energy (NREL), a US federal laboratory, identified a prom-

ising application of second-hand plug-in hybrid (PHEV) battery packs in grid energy
peak shaving.190 AI algorithms have been researched for the estimation of RUL14,191
and can be applied to estimate the usage of disposed batteries before recondition-
ing for second-hand applications.

Challenges and perspective for AI in EVCS and EV integration with the smart
grid
Most of the approaches discussed in this review for the optimal EVCS location and
energy management require a model for energy demand and supply. This is a com-
plex model with factors with high uncertainties, such as the charging patterns of EV
users, which can challenge the model empirically or mathematically. Moreover,
sometimes it is desirable to assign qualitative metrics for some factors when suffi-
cient quantitative information is missing, such as quality of air. ML algorithms can
be used in such cases but are challenged by the lack of sufficient EV user charging
data. As the EV penetration in the automotive industry increases, this EV user
charging data can become more accessible. CAN-Dataset is one of the publicly
available EVCS charging datasets, which has data points on charging sessions
from various on-site charging infrastructures.192 Moreover, the approach of using
RL and deep learning for load prediction in the smart grid has been deemed more
commercially viable than that using traditional ML techniques.193 It is also worth
noting that due to high investment, legal and political barriers, and the actual
construction and operation of EVCS at theoretically determined optimal EVCS site
location might face considerate obstacles and, therefore, hinder its actual imple-
mentation.194,195 Meanwhile, the discussed methods for improving and upgrading
the existing EVCS infrastructure might hold considerably higher commercial merit.

The development and maintenance of EVCSs and other EV-charging infrastructure

will have to keep on pace with the growing EV fleet.196 Apart from optimal site place-
ment of EVCS, ML provides the accommodation of EV-charging demand by the
existing EVCS infrastructure through EVCS congestion management and demand-
dependent pricing.157,197 By estimating the electric demand for the EVCS by using
ML models, utility and EVCS owners can dynamically change the pricing or charging
sessions to curb or encourage EV users and ensure EVCS economic viability.197
Recently, Valogianni et al. introduced adaptive pricing, whereby the prices are
adjusted based on EV users’ price reactions, and this pricing outperforms conven-
tional pricing schemes.198 ML models can be ideal candidates for future adaptive
pricing because of their ability to model the complex interaction between EV users’
behavior, external conditions, and price changes.

ACKNOWLEDGMENTS
The authors would like to acknowledge the financial support from the Natural Sci-
ences and Engineering Research Council of Canada (NSERC), Mitacs, the University
of Waterloo, and the Waterloo Institute for Nanotechnology.

AUTHOR CONTRIBUTIONS
M.A. and Y.Z. conceived the review article. M.A. conducted relevant literature re-
views and wrote the original manuscript. Y.Z. did major revisions and conducted
literature review accordingly. H.F. conducted relevant literature review and wrote
part of the section on battery-management systems. A.A edited and proofed the
manuscript. M.A., Y.Z., A.A., and Z.C. discussed, reviewed, and revised the manu-
script. Z.C. and Y.Z. provided supervision throughout the manuscript-writing

Joule 5, 2296–2322, September 15, 2021 2317

 ll
Review

process. M.A. and Y.Z. contributed equally to this work. All authors approved the
final version of the manuscript.

DECLARATION OF INTERESTS
The authors declare no competing interests.

REFERENCES
1. Perkins, J.H. (2017). Changing Energy: the 14. Hu, X., Xu, L., Lin, X., and Pecht, M. (2020). 28. Hanifah, R.A., Toha, S.F., Ahmad, S., and
Transition to a Sustainable Future (University Battery lifetime prognostics. Joule 4, 310–346. Hassan, M.K. (2018). Swarm-intelligence
of California Press). tuned current reduction for power-assisted
15. Grosan, C., and Abraham, A. (2011). Rule- steering control in electric vehicles. IEEE
2. Energy flow LlNLcharts (2021). https:// based expert systems. In Intelligent Systems: Trans. Ind. Electron. 65, 7202–7210.
flowcharts.llnl.gov/. A Modern Approach (Springer), pp. 149–185.
29. Soares, J., Vale, Z., Canizes, B., and Morais, H.
3. UNFCCC The Paris Agreement (2018). 16. Geron, A. (2019). The machine learning (2013). Multi-objective parallel particle swarm
https://unfccc.int/process-and-meetings/ landscape. In Hands Machine Learning with optimization for day-ahead vehicle-to-grid
the-paris-agreement/the-paris-agreement. Scikit-Learn and Tensorflow (O’Reilly), scheduling. In IEEE Symposium on
pp. 7–17. Computational Intelligence Applications in
4. Bilgin, B., Magne, P., Malysz, P., Yang, Y., Smart Grid (CIASG), pp. 138–145.
17. Chen, A., Zhang, X., and Zhou, Z. (2020).
Pantelic, V., Preindl, M., Korobkine, A., Jiang,
Machine learning: accelerating materials 30. Gopalakrishnan, R., Biswas, A., Lightwala, A.,
W., Lawford, M., and Emadi, A. (2015). Making
development for energy storage and Vasudevan, S., Dutta, P., and Tripathi, A.
the case for electrified transportation. IEEE
conversion. InfoMat 2, 553–576. (2016). Demand prediction and placement
Trans. Transp. Electrific. 1, 4–17.
optimization for electric vehicle charging
18. Cunha, R.P., Lombardo, T., Primo, E.N., and stations. In IJCAI International Joint
5. International Energy Agency (2020). Global Franco, A.A. (2020). Artificial intelligence
EV Outlook 2020. https://www.iea.org/ Conference on Artificial Intelligence,
investigation of NMC cathode manufacturing pp. 3117–3123.
reports/global-ev-outlook-2020. parameters interdependencies. Batteries
Supercaps 3, 60–67. 31. Zheng, Y., Yao, Y., Ou, J., Li, M., Luo, D., Dou,
6. Sathe, P. (2020). Global Electric Vehicle
H., Li, Z., Amine, K., Yu, A., and Chen, Z.
Market Outlook, 2020. Global Electric Vehicle 19. Varga, B., Sagoian, A., and Mariasiu, F. (2019). (2020). A review of composite solid-state
Market Outlook (Frost & Sullivan). https:// Prediction of electric vehicle range: a electrolytes for lithium batteries:
store.frost.com/global-electric-vehicle- comprehensive review of current issues and fundamentals, key materials and advanced
market-outlook-2020.html. challenges. Energies 12, 946. structures. Chem. Soc. Rev. 49, 8790–8839.
7. BloombergNEF. (2019). Battery pack prices 20. Abu Hanifah, R., Toha, S.F., Hassan, M.K., and 32. Alkheir, A.A., Aloqaily, M., and Mouftah, H.T.
fall as market ramps up with market average Ahmad, S. (2016). Power reduction (2018). Connected and autonomous electric
at $156/kWh in 2019. https://about.bnef.com/ optimization with swarm based technique in vehicles (CAEVs). IT Prof 20, 54–61.
blog/battery-pack-prices-fall-as-market- electric power assist steering system. Energy
ramps-up-with-market-average-at-156-kwh- 102, 444–452. 33. Mitropoulos, L.K., Prevedouros, P.D., and
in-2019/. Kopelias, P. (2017). Total cost of ownership
21. Li, J., Cheng, H., Guo, H., and Qiu, S. (2018). and externalities of conventional, hybrid and
8. McKinsey & Company. (2019). The future of Correction to: Survey on artificial intelligence electric vehicle. Transp. Res. Procedia 24,
mobility is at our doorstep. https://www. for vehicles. Automot. Innov. 1, 390. 267–274.
mckinsey.com//media/McKinsey/
Industries/Automotive%20and%20Assembly/ 22. Rigas, E.S., Ramchurn, S.D., and Bassiliades, 34. Barrett, D.H., and Haruna, A. (2020). Artificial
Our%20Insights/The%20future%20of% N. (2015). Managing electric vehicles in the intelligence and machine learning for
20mobility%20is%20at%20our%20doorstep/ smart grid using artificial intelligence: a targeted energy storage solutions. Curr.
The-future-of-mobility-is-at-our-doorstep. survey. In IEEE Transactions on Intelligent Opin. Electrochem. 21, 160–166.
ashx. Transportation Systems, 16, pp. 1619–1635.
35. Choi, H., Sohn, K.S., Pyo, M., Chung, K.C., and
9. Un-Noor, F., Padmanaban, S., Mihet-Popa, L., 23. Hannan, M.A., Lipu, M.S.H., Hussain, A., Ker, Park, H. (2019). Predicting the electrochemical
Mollah, M.N., and Hossain, E. (2017). A P.J., Mahlia, T.M.I., Mansor, M., Ayob, A., properties of lithium-ion battery electrode
comprehensive study of key electric vehicle Saad, M.H., and Dong, Z.Y. (2020). Toward materials with the quantum neural network
(EV) components, technologies, challenges, enhanced state of charge estimation of algorithm. J. Phys. Chem. C 123, 4682–4690.
impacts, and future direction of lithium-ion batteries using optimized
development. Energies 10, 1217. machine learning techniques. Sci. Rep. 10, 36. Heng, K.L., Jin, H.M., Li, Y., and Wu, P. (1999).
4687. Computer aided design of NiMH electrodes.
J. Mater. Chem. 9, 837–843.
10. Neubauer, J., and Wood, E. (2014). The
24. Chen, C., Zuo, Y., Ye, W., Li, X., Deng, Z., and
impact of range anxiety and home, 37. Min, K., Choi, B., Park, K., and Cho, E. (2018).
Ong, S.P. (2020). A critical review of machine
workplace, and public charging infrastructure Machine learning assisted optimization of
learning of energy materials. Adv. Energy
on simulated battery electric vehicle lifetime
Mater. 10, 1903242. electrochemical properties for Ni-rich
utility. J. Power Sources 257, 12–20. cathode materials. Sci. Rep. 8, 15778.
25. Geron, A. (2019). Introduction to artificial
11. Winston, P. (1992). Chapter 1. The intelligent neural networks with keras. In Hands Machine 38. Attarian Shandiz, M., and Gauvin, R. (2016).
computer. In Artificial Intelligence (Wesley Learning with Scikit-Learn and Tensorflow, Application of machine learning methods for
Longman Publishing). pp. 280–281. the prediction of crystal system of cathode
materials in lithium-ion batteries.
12. Ng, M., Zhao, J., Yan, Q., Conduit, G.J., and 26. Geron, A. (2019). Reinforcement learning. In Computational Materials Science 117,
Seh, Z.W. (2020). Let’s go 2020. Nat. Mach. Hands Machine Learning with Scikit-Learn 270–278.
Intell. 2, 1–10. and Tensorflow (O’Reilly), pp. 609–664.
39. Fujimura, K., Seko, A., Koyama, Y., Kuwabara,
13. Grigorescu, S., Trasnea, B., Cocias, T., and 27. Slowik, A., and Kwasnicka, H. (2020). A., Kishida, I., Shitara, K., Fisher, C.A.J.,
Macesanu, G. (2019). A survey of deep Evolutionary algorithms and their Moriwake, H., and Tanaka, I. (2013).
learning techniques for autonomous driving. applications to engineering problems. Neural Accelerated materials design of lithium
J. Field Robotics 37, 362–386. Comput. & Applic. 32, 12363–12379. superionic conductors based on first-

2318 Joule 5, 2296–2322, September 15, 2021

 ll
Review

principles calculations and machine learning Herring, P.K., Fraggedakis, D., et al. (2019). design of lithium-ion batteries. J. Power
algorithms. Adv. Energy Mater. 3, 980–985. Data-driven prediction of battery cycle life Sources 395, 128–136.
before capacity degradation. Nat. Energy 4,
40. Sendek, A.D., Cubuk, E.D., Antoniuk, E.R., 383–391. 66. Cheng, K.W.E., Divakar, B.P., Wu, H., Ding, K.,
Cheon, G., Cui, Y., and Reed, E.J. (2019). and Ho, H.F. (2011). Battery-management
Machine learning-assisted discovery of solid 54. Attia, P.M., Grover, A., Jin, N., Severson, K.A., system (BMS) and SOC development for
Li-ion conducting materials. Chem. Mater. 31, Markov, T.M., Liao, Y.H., Chen, M.H., Cheong, electrical vehicles. IEEE Trans. Veh. Technol.
342–352. B., Perkins, N., Yang, Z., et al. (2020). Closed- 60, 76–88.
loop optimization of fast-charging protocols
41. Okamoto, Y., and Kubo, Y. (2018). Ab initio for batteries with machine learning. Nature 67. Xiong, R., Cao, J., Yu, Q., He, H., and Sun, F.
calculations of the redox potentials of 578, 397–402. (2017). Critical review on the battery state of
additives for lithium-ion batteries and their charge estimation methods for electric
prediction through machine learning. ACS 55. Gayon-Lombardo, A., Mosser, L., Brandon, vehicles. IEEE Access 6, 1832–1843.
Omega 3, 7868–7874. N.P., and Cooper, S.J. (2020). Pores for
thought: generative adversarial networks for 68. Parakkadavath, S., and Bhikkaji, B. (2017).
42. Tkatchenko, A. (2020). Machine learning for stochastic reconstruction of 3D multi-phase Identification of the non-linear dynamics and
chemical discovery. Nat. Commun. 11, 4125. electrode microstructures with periodic state of charge estimation of a LiFePO4
boundaries. npj Comput. Mater. 6, 1–11. battery using constrained unscented Kalman
43. Jain, A., Ong, S.P., Hautier, G., Chen, W., filter. IFAC-PapersOnLine 50, 1571–1576.
Richards, W.D., Dacek, S., et al. (2013). 56. Ding, Y., Cano, Z.P., Yu, A., Lu, J., and Chen, Z.
Commentary: The Materials Project: A (2019). Automotive Li-ion batteries: current 69. Vidal, C., Malysz, P., Kollmeyer, P., and Emadi,
materials genome approach to accelerating status and future perspectives. Electrochem. A. (2020). Machine learning applied to
materials innovation. APL Mater 1, 011002. Energ. Rev. 2, 1–28. electrified vehicle battery state of charge and
state of health estimation: state-of-the-art.
44. Reymond, J.L. (2015). The chemical space 57. Himanen, L., Geurts, A., Foster, A.S., and IEEE Access 8, 52796–52814.
project. Acc. Chem. Res. 48, 722–730. Rinke, P. (2019). Data-driven materials
science: status, challenges, and perspectives. 70. Feng, F., Teng, S., Liu, K., Xie, J., Xie, Y., Liu,
45. Allam, O., Cho, B.W., Kim, K.C., and Jang, S.S. B., and Li, K. (2020). Co-estimation of lithium-
(2018). Application of DFT-based machine Adv. Sci. (Weinh) 6, 1900808.
ion battery state of charge and state of
learning for developing molecular electrode temperature based on a hybrid
58. Meredig, B. (2017). Industrial materials
materials in Li-ion batteries. RSC Adv. 8, electrochemical-thermal-neural-network
informatics: analyzing large-scale data to
39414–39420. model. J. Power Sources 455, 227935.
solve applied problems in R&D,
46. Hautier, G., Fischer, C.C., Jain, A., Mueller, T., manufacturing, and supply chain. Curr. Opin.
71. Karlsen, H., Dong, T., Yang, Z., and Carvalho,
and Ceder, G. (2010). Finding nature’s missing Solid State Mater. Sci. 21, 159–166.
R. (2019). Temperature-dependence in
ternary oxide compounds using machine battery management systems for electric
59. Ahmed, M., Yazdi, A.Z., Dayani, S.B., Jahed,
learning and density functional theory. Chem. vehicles: challenges, criteria, and solutions.
H., and Chen, P. (2019). Fabrication of zinc
Mater. 22, 3762–3767. IEEE Access 7, 142203–142213.
anodes for aqueous lithium-ion batteries by
47. Wu, X., Kang, F., Duan, W., and Li, J. (2019). supersonic cold spraying. ChemElectroChem
72. Ojo, O., Lang, H., Kim, Y., Hu, X., Mu, B., and
Density functional theory calculations: a 6, 1333–1337.
Lin, X. (2021). A neural network based method
powerful tool to simulate and design high- for thermal fault detection in lithium-ion
60. Takagishi, Y., Yamanaka, T., and Yamaue, T.
performance energy storage and conversion batteries. IEEE Trans. Ind. Electron. 68, 4068–
(2019). Machine learning approaches for
materials. Prog. Nat. Sci. Mater. Int. 29, 4078.
designing mesoscale structure of Li-ion
247–255.
battery electrodes. Batteries 5, 54.
73. Chen, C.-H., Brosa Planella, F., O’Regan, K.,
48. Liu, K., Wei, Z., Yang, Z., and Li, K. (2021). Mass Gastol, D., Widanage, W.D., and Kendrick, E.
load prediction for lithium-ion battery 61. Ruhatiya, C., Singh, S., Goyal, A., Niu, X., Hanh
Nguyen, T.N., Nguyen, V.H., Tran, V.M., (2020). Development of experimental
electrode clean production: a machine techniques for parameterization of multi-
learning approach. J. Clean. Prod. 289, Phung LE, M.L., Garg, A., and Gao, L. (2020).
Electrochemical performance enhancement scale lithium-ion battery models.
125159. J. Electrochem. Soc. 167, 080534.
of sodium-ion batteries fabricated with
49. Duquesnoy, M., Lombardo, T., Chouchane, NaNi1/3Mn1/3Co1/3O2 cathodes using 74. Center for Advanced Life Cycle Engineering
M., Primo, E.N., and Franco, A.A. (2020). Data- support vector regression-simplex algorithm Battery data. (University of Maryland) (2021).
driven assessment of electrode calendering approach. J. Electrochem. Energy Convers. https://calce.umd.edu/battery-data.
process by combining experimental results, Storage 17.
in silico mesostructures generation and 75. Charkhgard, M., and Farrokhi, M. (2010).
machine learning. J. Power Sources 480, 62. Homma, K., Liu, Y., Sumita, M., Tamura, R., State-of-charge estimation for lithium-ion
229103. Fushimi, N., Iwata, J., Tsuda, K., and Kaneta, batteries using neural networks and EKF. IEEE
C. (2020). Optimization of a heterogeneous Trans. Ind. Electron. 57, 4178–4187.
50. Buteau, S., and Dahn, J.R.D.J. (2019). Analysis ternary Li3 PO4 –Li3 BO3 –Li2 SO4 mixture for
of thousands of electrochemical impedance Li-ion conductivity by machine learning. 76. Nuhic, A., Terzimehic, T., Soczka-Guth, T.,
spectra of lithium-ion cells through a machine J. Phys. Chem. C 124, 12865–12870. Buchholz, M., and Dietmayer, K. (2013). Health
learning inverse model. J. Electrochem. Soc. diagnosis and remaining useful life
166, A1611–A1622. 63. Ruhatiya, C., Singh, S., Goyal, A., Niu, X., prognostics of lithium-ion batteries using
Ngoc, T., Nguyen, H., Nguyen, V.H., Tran, data-driven methods. J. Power Sources 239,
51. Ellis, L.D., Buteau, S., Hames, S.G., V.M., Loan, M., Le, P., et al. (2020). 680–688.
Thompson, L.M., Hall, D.S., and Dahn, J.R. Electrochemical performance enhancement
(2018). A new method for determining the of sodium-ion batteries fabricated with using 77. Yang, F., Li, W., Li, C., and Miao, Q. (2019).
concentration of electrolyte components in support vector regression-simplex algorithm State-of-charge estimation of lithium-ion
lithium-ion cells, using Fourier transform approach. J. Electrochem. En. Conv. Stor. 17, batteries based on gated recurrent neural
infrared spectroscopy and machine learning. 1–8. network. Energy 175, 66–75.
J. Electrochem. Soc. 165, A256–A262.
64. Li, Y., Liu, K., Foley, A.M., Zülke, A., Berecibar, 78. Venugopal, P., and T., V. (2019). State-of-
52. Zhu, S., Sun, X., Gao, X., Wang, J., Zhao, N., M., Nanini-maury, E., Van Mierlo, J., and health estimation of Li-ion batteries in electric
and Sha, J. (2019). Equivalent circuit model Hoster, H.E. (2019). Data-driven health vehicle using IndRNN under variable load
recognition of electrochemical impedance estimation and lifetime prediction of lithium- condition. Energies 12, 4338.
spectroscopy via machine learning. ion batteries: a review. Renewable and
J. Electroanal. Chem. 855, 113627. Sustainable Energy Reviews 113, 109254. 79. Hashemi, S.R., Bahadoran Baghbadorani, A.,
Esmaeeli, R., Mahajan, A., and Farhad, S.
53. Severson, K.A., Attia, P.M., Jin, N., Perkins, N., 65. Wu, B., Han, S., Shin, K.G., and Lu, W. (2018). (2021). Machine learning-based model for
Jiang, B., Yang, Z., Chen, M.H., Aykol, M., Application of artificial neural networks in lithium-ion batteries in BMS of electric/hybrid

Joule 5, 2296–2322, September 15, 2021 2319

 ll
Review

electric aircraft. Int. J. Energy Res. 45, 5747– vehicle driving range. Resour. Energy Econ 108. Nian, Xiaohong, Peng, Fei, and Zhang, Hang
5765. 35, 429–450. (2014). Regenerative braking system of
electric vehicle driven by brushless DC motor.
80. Deng, J., Bae, C., Marcicki, J., Masias, A., and 95. Çeven, S., Albayrak, A., and Bayır, R. (2020). IEEE Trans. Ind. Electron. 61, 5798–5808.
Miller, T. (2018). Safety modelling and testing Real-time range estimation in electric vehicles
of lithium-ion batteries in electrified vehicles. using fuzzy logic classifier. Comput. Electr. 109. Doyle, A., and Muneer, T. (2017). Traction
Nat. Energy 3, 261–266. Eng. 83, 106577. energy and battery performance modelling.
In Electric Vehicles: Prospects and Challenges
81. Yao, L., Xiao, Y., Gong, X., Hou, J., and Chen, 96. Zhang, Y., Wang, W., Kobayashi, Y., and (Elsevier), pp. 93–124.
X. (2020). A novel intelligent method for fault Shirai, K. (2012). Remaining driving range
diagnosis of electric vehicle battery system estimation of electric vehicle. In IEEE 110. Gupta, R.A., Kumar, R., and Bansal, A.K.
based on wavelet neural network. J. Power International Electric Vehicle Conference (2010). Artificial intelligence applications in
Sources 453, 227870. (IEVC 2012), pp. 1–7. Permanent Magnet Brushless DC motor
drives. Artif. Intell. Rev. 33, 175–186.
82. Ojo, O., Lang, H., Kim, Y., Hu, X., Mu, B., and 97. Pan, C., Dai, W., Chen, L., Chen, L., and Wang,
Lin, X. (2020). A neural network-based method L. (2017). Driving range estimation for electric 111. Abulifa, A.A., Che Soh, A., Hassan, M.K., Raja
for thermal fault detection in lithium-ion vehicles based on driving condition Ahmad, R.M.K., and Mohd Radzi, M.A. (2019).
batteries. IEEE Trans. Ind. Electron. 68, 4068– identification and forecast. AIP Adv. 7, Energy management system in battery
4078. 105206. electric vehicle based on fuzzy logic control to
optimize the energy consumption in HVAC
83. Li, W., Chen, S., Peng, X., Xiao, M., Gao, L., 98. Sun, S., Zhang, J., Bi, J., Wang, Y., and system. Int. J. Integr. Eng. 11, 11–20.
Garg, A., and Bao, N. (2019). A Moghaddam, M.H.Y. (2019). A machine
comprehensive approach for the clustering of 112. Chu, L., Yao, L., Yin, J., Chao, L., and Wei, W.
learning method for predicting driving range
similar-performance cells for the design of a (2011). Study on the braking force allocation
of battery electric vehicles. J. Adv. Transp.
lithium-ion battery module for electric dynamic control strategy based on the fuzzy
2019, 1–14.
vehicles. Engineering 5, 795–802. control logic. In 2011 IEEE 18th International
Conference on Industrial Engineering and
99. Yavasoglu, H.A., Tetik, Y.E., and Gokce, K.
84. Yao, L., Wang, Z., and Ma, J. (2015). Fault Engineering Management, pp. 635–639.
(2019). Implementation of machine learning
detection of the connection of lithium-ion based real time range estimation method
power batteries based on entropy for electric 113. Zhang, H., Xu, G., Li, W., and Zhou, M. (2012).
without destination knowledge for BEVs. Fuzzy logic control in regenerative braking
vehicles. J. Power Sources 293, 548–561. Energy 172, 1179–1186. system for electric vehicle. In IEEE
85. Jia, Y., Liu, B., Hong, Z., Yin, S., Finegan, D.P., International Conference on Information and
100. Zheng, B., He, P., Zhao, L., and Li, H. (2016). A Automation (ICIA 2012), pp. 588–591.
and Xu, J. (2020). Safety issues of defective hybrid machine learning model for range
lithium-ion batteries: identification and risk estimation of electric vehicles. In Proceedings 114. Xu, G., Li, W., Xu, K., and Song, Z. (2011). An
evaluation. J. Mater. Chem. A 8, 12472–12484. IEEE Global Communications Conference, intelligent regenerative braking strategy for
GLOBECOM 2016 (Institute of Electrical and electric vehicles. Energies 4, 1461–1477.
86. Bishop, W.S. (1984). Battery control and fault
Electronics Engineers), pp. 1–6.
detection method (United States Air Force). 115. Jianbo, C., Binggang, C., Wenzhi, C., and
https://patents.google.com/patent/ 101. Bolovinou, A., Bakas, I., Amditis, A., Peng, X. (2007). Neural network self-adaptive
US4633418. Mastrandrea, F., and Vinciotti, W. (2014). PID control for driving and regenerative
Online prediction of an electric vehicle braking of electric vehicle. In Proceedings of
87. Keates, A.W., Otani, N., Nguyen, D.J.,
remaining range based on regression the IEEE International Conference on
Matsumura, N., and Li, P.T. (2010). Short
analysis. In IEEE International Electric Vehicle Automation and Logistics (ICAL 2007),
circuit detection for batteries (Intel Corp.).
Conference (IEVC 2014), pp. 1–8. pp. 2029–2034.
https://patents.google.com/patent/
US7795843B2/en. 102. Scheubner, S., Thorgeirsson, A.T., Vaillant, M., 116. He, H., Wang, C., Jia, H., and Cui, X. (2020). An
and Gauterin, F. (2019). A stochastic range intelligent braking system composed single-
88. Sivertsen, D. (2009). Detecting faults in a
estimation algorithm for electric vehicles pedal and multi-objective optimization neural
wiring harness. US Patent, US 8,164,433 B2,
using traffic phase classification. IEEE Trans. network braking control strategies for electric
filed April 17, 2009 and granted April 24, 2012.
Veh. Technol. 68, 6414–6428. vehicle. Appl. Energy 259, 114172.
89. Judge, A., Renehan, J., and Deljevic, Z. (1999). 117. Johnson, M.A. (2005). PID control technology.
Fault detection in a motor vehicle charging 103. Rhode, S., Van Vaerenbergh, S., and Pfriem,
M. (2020). Power prediction for electric In PiD Control (Springer-Verlag), pp. 1–46.
system. US Patent, US 6,194,877 9 9 B1, filed
August 2, 1999 and granted February 27, vehicles using online machine learning. Eng.
118. Ekinci, S., Hekimoǧlu, B., and Kaya, S. (2019).
2001. Appl. Artif. Intell. 87, 103278.
Tuning of PID controller for AVR system using
104. Yokoi, Y., Ichikawa, S., Doki, S., Okuma, S., salp swarm algorithm. In International
90. Liu, B., Jia, Y., Yuan, C., Wang, L., Gao, X., Yin, Conference on Artificial Intelligence and Data
S., and Xu, J. (2020). Safety issues and Naitou, T., Shiimado, T., and Miki, N. (2004).
Driving pattern prediction for an energy Processing (IDAP 2018 (Institute of Electrical
mechanisms of lithium-ion battery cell upon and Electronics Engineers)), pp. 1–6.
mechanical abusive loading: a review. Energy management system of hybrid electric
Storage Mater 24, 85–112. vehicles in a specific driving course. In IECON 119. Wati, D.A.R. (2013). Performance evaluation of
Proceedings (Industrial Electronics swarm intelligence on model-based PID
91. Li, W., Zhu, J., Xia, Y., Gorji, M.B., and Conference), pp. 1727–1732. tuning. In Proceeding - IEEE
Wierzbicki, T. (2019). Data-driven safety CYBERNETICSCOM 2013: IEEE International
envelope of lithium-ion batteries for electric 105. Lee, C.H., and Wu, C.H. (2015). A novel big
Conference on Computational Intelligence
vehicles. Joule 3, 2703–2715. data modeling method for improving driving
and Cybernetics, pp. 40–44.
range estimation of EVs. IEEE Access 3, 1980–
92. Finegan, D.P., and Cooper, S.J. (2019). Battery 1993. 120. Li, X., Wen, H., Hu, Y., and Jiang, L. (2019). A
safety: data-driven prediction of failure. Joule novel beta parameter based fuzzy-logic
3, 2599–2601. 106. Nowaková, J., and Pokorný, M. (2020). controller for photovoltaic MPPT application.
Intelligent controller design by the artificial Renew. Energy 130, 416–427.
93. Finegan, D.P., Zhu, J., Feng, X., Keyser, M., intelligence methods. Sensors (Basel) 20,
Ulmefors, M., Li, W., Bazant, M.Z., and 1–27. 121. Mirinejad, H., Welch, K.C., and Spicer, L.
Cooper, S.J. (2021). The application of data- (2012). A review of intelligent control
driven methods and physics-based learning 107. Wicaksono, A., and Prihatmanto, A.S. (2016). techniques in HVAC systems. In IEEE
for improving battery safety. Joule 5, 316–329. Optimal control system design for electric Energytech Energytech 2012.
vehicle. In Proceedings of the 2015 4th
94. Daziano, R.A. (2013). Conditional-logit Bayes International Conference on Interactive 122. Tsai, C.C., Hwang, K.S., Liu, A., and Juang,
estimators for consumer valuation of electric Digital Media (ICIDM 2015), pp. 1–6. C.F. (2015). Advances and challenges on

2320 Joule 5, 2296–2322, September 15, 2021

 ll
Review

intelligent learning in control systems. In 135. Xu, J., Li, J., Liao, X., and Song, C. (2019). A Victorian EV Trial, Australia. Transp. Res. D 32,
Contemporary Issues in Systems Science and prediction method of charging station 263–277.
Engineering (Wiley-IEEE Press), pp. 254–255. planning based on BP neural network.
J. Comput. Commun. 07, 219–230. 150. Hu, J., Morais, H., Lind, M., and Bindner, H.W.
123. Srdic, S., and Lukic, S. (2019). Toward extreme (2016). Multi-agent based modeling for
fast charging: challenges and opportunities in  zlavský, O. (2014). Net present value
136. Zi electric vehicle integration in a distribution
directly connecting to medium-voltage line. approach: method for economic assessment network operation. Electr. Power Syst. Res.
IEEE Electrific Mag. 7, 22–31. of innovation projects. Procedia-Soc. Behav. 136, 341–351.
Sci. 156, 506–512.
124. Dubey, A., and Santoso, S. (2015). Electric 151. Miranda, J., Borges, J., Mendes, M.J.G.C.,
vehicle charging on residential distribution 137. Andrenacci, N., Ragona, R., and Valenti, G. and Valério, D. (2011). Development of a
systems: impacts and mitigations. IEEE (2016). A demand-side approach to the multi-agent management system for an
Access 3, 1871–1893. optimal deployment of electric vehicle intelligent charging network of electric
charging stations in metropolitan areas. Appl. vehicles. IFAC Proceedings Volumes 44,
125. Elaadnl. (2021). Data analytics. https://www. Energy 182, 39–46. 12267–12272.
elaad.nl/research/data-analytics/.
138. Hosseini, S., and Sarder, M.D. (2019). 152. Xydas, E., Marmaras, C., Cipcigan, L.M.,
126. Ip, A., Fong, S., and Liu, E. (2010). Development of a Bayesian network model Jenkins, N., Carroll, S., and Barker, M. (2016).
Optimization for allocating BEV recharging for optimal site selection of electric vehicle A data-driven approach for characterising the
stations in urban areas by using hierarchical charging station. Int. J. Electr. Power Energy charging demand of electric vehicles: a UK
clustering. In Proc. - 6th Intl. Conference on Syst. 105, 110–122. case study. Appl. Energy 162, 763–771.
Advanced Information Management and 153. Ramachandran, A., Balakrishna, A., Kundzicz,
139. Wolbertus, R., van den Hoed, R., Kroesen, M.,
Service, IMS2010, with ICMIA2010 - 2nd P., and Neti, A. (2018). Predicting electric
and Chorus, C. (2021). Charging infrastructure
International Conference on Data Mining and vehicle charging station usage: using machine
roll-out strategies for large scale introduction
Intelligent Information Technology learning to estimate individual station
of electric vehicles in urban areas: an agent-
Applications (Intl. Conference on Advanced
based simulation study. Transp. Res. A 148, statistics from physical configurations of
Information Management and Service), charging station networks. arXiv,
262–285.
pp. 1–6. arXiv:1804.00714.
140. Sweda, T., and Klabjan, D. (2011). An agent-
127. Liu, Z.F., Zhang, W., Ji, X., and Li, K. (2012). based decision support system for electric 154. Chung, Y.W., Khaki, B., Li, T., Chu, C., and
Optimal planning of charging station for vehicle charging infrastructure deployment. In Gadh, R. (2019). Ensemble machine learning-
electric vehicle based on particle swarm IEEE Vehicle Power and Propulsion based algorithm for electric vehicle user
optimization. In IEEE Innovative Smart Grid Conference, VPPC 2011, pp. 1–5. behavior prediction. Appl. Energy 254,
Technologies - Asia (ISGT 2012), pp. 1–5. 113732.
141. Sheppard, C.J.R., Harris, A., and Gopal, A.R.
128. Wu, H., and Niu, D. (2017). Study on factors of (2016). Cost-effective siting of electric vehicle 155. Zhu, J., Yang, Z., Mourshed, M., Guo, Y., Zhou,
electric vehicles charging station location charging infrastructure with agent-based Y., Chang, Y., Wei, Y., and Feng, S. (2019).
based on ISM and FMICMAC. Sustainability 9, modeling. IEEE Trans. Transp. Electrific. 2, Electric vehicle charging load forecasting: A
48. 174–189. comparative study of deep learning
approaches. Energies 12.
129. Pevec, D., Babic, J., Kayser, M.A., Carvalho, 142. O’Sullivan, D., Millington, J., Perry, G., and
A., Ghiassi-Farrokhfal, Y., and Podobnik, V. Wainwright, J. (2012). Agent-based models- 156. Frendo, O., Graf, J., Gaertner, N., and
(2018). A data-driven statistical approach for because they’re worth it? In Agent-Based Stuckenschmidt, H. (2020). Data-driven smart
extending electric vehicle charging Models of Geographical Systems (Springer charging for heterogeneous electric vehicle
infrastructure. Int. J. Energy Res. 42, 3102– Netherlands), pp. 109–123. fleets. Energy and AI 1, 100007.
3120.
143. Dahlke, J., Bogner, K., Mueller, M., Berger, T., 157. Zhang, L., Gong, K., and Xu, M. (2019).
130. Li, Y., Li, L., Yong, J., Yao, Y., and Li, Z. (2011). Pyka, A., and Ebersberger, B. (2020). Is the Congestion control in charging stations
Layout planning of electrical vehicle charging Juice Worth the Squeeze? Machine Learning allocation with Q-learning. Sustainability 11,
stations. In Electrical Power Systems and (ML) in and for Agent-Based Modelling 1–11.
Computers (Springer), pp. 661–668. (ABM). arXiv, arXiv:2003.11985.
158. Sherstyuk, M., Sherstyuk, T., and Prokoptsov,
131. Deb, S., Kalita, K., Gao, X.Z., Tammi, K., and 144. Kavak, H., Padilla, J.J., Lynch, C.J., and Diallo, M. (2018). Systems and methods for
Mahanta, P. (2017). Optimal placement of S.Y. (2018). Big Data, Agents, and Machine enhancing the performance and utilization of
charging stations using CSO-TLBO algorithm. Learning: Towards a Data-Driven Agent- battery systems (G Batteries Energy Inc.).
In Proceedings 3rd IEEE International Based Modeling Approach. In ANSS ’18: https://patents.google.com/patent/
Conference on Research in Computational Proceedings of the Annual Simulation US10218200B2/en.
Intelligence and Communication Networks Symposium, pp. 125–136.
159. Tkachenko, O., Sherstyuk, M., Prodic, A.,
(ICRCICN 2017), pp. 84–89.
145. Cross, J.D., and Hartshorn, R. (2016). My McRae, T., Radovic, I., and Puzakov, O. (2018).
electric avenue: integrating electric vehicles Modulated pulse charging and discharging of
132. Awasthi, A., Venkitusamy, K., Padmanaban, S.,
into the electrical networks. In 6th Hybrid and a reconfigurable battery pack (GBatteries
Selvamuthukumaran, R., Blaabjerg, F., and
Electric Vehicles Conference (HEVC 2016), Energy Canada). https://patents.google.
Singh, A.K. (2017). Optimal planning of
pp. 1–6. com/patent/US20180219390A1/en.
electric vehicle charging station at the
distribution system using hybrid optimization 160. Tkachenko, O., Sherstyuk, M., Prodic, A.,
146. About dataport. (2021). Pecan Street. https://
algorithm. Energy 133, 70–78. Mcrae, T., Radovic, I., and Puzakov, O. (2020).
www.pecanstreet.org/dataport/about/.
Battery charging through multi-stage voltage
133. Deb, S., Gao, X., Tammi, K., Kalita, K., and 147. Shahriar, S., Al-Ali, A.R., Osman, A.H., Dhou, conversion.
Mahanta, P. (2021). Nature-Inspired S., and Nijim, M. (2020). Machine learning
Optimization Algorithms Applied for Solving approaches for EV charging behavior: a 161. Tomaszewska, A., Chu, Z., Feng, X., O’Kane,
Charging Station Placement Problem: review. IEEE Access 8, 168980–168993. S., Liu, X., Chen, J., Ji, C., Endler, E., Li, R., Liu,
Overview and Comparison. Arch. Computat. L., et al. (2019). Lithium-Ion Battery Fast
Methods Eng. 28, 91–106. 148. Angione, C., Silverman, E., and Yaneske, E. Charging: A Review. eTransportation 1,
(2020). Using machine learning to emulate 100011.
134. Mohsenzadeh, A., Pang, C., Pazouki, S., and agent-based simulations. arXiv,
Haghifam, M. (2015). Optimal siting and sizing arXiv:2005.02077v2. 162. Min, H., Sun, W., Li, X., Guo, D., Yu, Y., Zhu, T.,
of electric vehicle public charging stations and Zhao, Z. (2017). Research on the optimal
considering smart distribution network 149. Khoo, Y.B., Wang, C.H., Paevere, P., and charging strategy for Li-ion batteries based
reliability. In 2015 North American Power Higgins, A. (2014). Statistical modeling of on multi-objective optimization. Energies 10,
Symposium (NAPS), pp. 1–6. electric vehicle electricity consumption in the 709.

Joule 5, 2296–2322, September 15, 2021 2321

 ll
Review

163. Liu, Y.-H., Teng, J.-H., and Lin, Y.-C. (2005). techniques. Renew. Sustain. Energy Rev. 53, twin for battery systems: cloud battery
Search for an optimal rapid charging pattern 720–732. management system with online state-of-
for lithium-ion batteries using ant colony charge and state-of-health estimation.
system algorithm. IEEE Trans. Ind. Electron. 176. Wang, Z., and Wang, S. (2013). Grid power J. Energy Storage 30, 101557.
52, 1328–1336. peak shaving and valley filling using vehicle-
to-grid systems. IEEE Trans. Power Delivery 188. Taeihagh, A., and Lim, H.S.M. (2019).
164. Surmann, H. (1996). Genetic optimization of a 28, 1822–1829. Governing autonomous vehicles: emerging
fuzzy system for charging batteries. IEEE responses for safety, liability, privacy,
Trans. Ind. Electron. 43, 541–548. 177. Mwasilu, F., Justo, J.J., Kim, E.-K., Do, T.D.,
cybersecurity, and industry risks. Transp. Rev.
and Jung, J.-W. (2014). Electric vehicles and
39, 103–128.
165. Garcı́a Álvarez, J., González, M., Rodrı́guez smart grid interaction: a review on vehicle to
Vela, C., and Varela, R. (2018). Electric vehicle grid and renewable energy sources
189. Harper, G., Sommerville, R., Kendrick, E.,
charging scheduling by an enhanced artificial integration. Renew. Sustain. Energy Rev. 34,
Driscoll, L., Slater, P., Stolkin, R., Walton, A.,
bee colony algorithm. Energies 11, 1–19. 501–516.
Christensen, P., Heidrich, O., Lambert, S.,
166. Wang, S.C., and Liu, Y.H. (2015). A PSO-based 178. Frost&Sullivan. (2017). Developments in et al. (2019). Recycling lithium-ion batteries
fuzzy-controlled searching for the optimal vehicle-to-grid (V2G) Technology, https:// from electric vehicles. Nature 575, 75–86.
charge pattern of Li-ion batteries. IEEE Trans. store.frost.com/developments-in-vehicle-to-
Ind. Electron. 62, 2983–2993. grid-v2g-technology.html. 190. Neubauer, J., Smith, K., Wood, E., and
Pesaran, A. (2015). Identifying and
167. Venayagamoorthy, G.K. (2011). Dynamic, 179. Noel, L., Zarazua de Rubens, G., Kester, J., overcoming critical barriers to widespread
stochastic, computational, and scalable Sovacool, B.K., Noel, L., Zarazua de Rubens, second use of PEV batteries.
technologies for smart grids. IEEE Comput. G., Kester, J., and Sovacool, B.K. (2019). The
Intell. Mag. 6, 22–35. technical challenges to V2G. In Vehicle Grid 191. Garg, A., Wei, L., Goyal, A., Cui, X., and Gao,
(Springer Int. Publ.), pp. 65–89. L. (2019). Evaluation of batteries residual
168. Frost&Sullivan. (2017). Smart grids in Eurpoe energy for battery pack recycling: proposition
and APAC. https://store.frost.com/smart- 180. Liu, D., Wang, Y., and Shen, Y. (2016). Electric of stack stress-coupled-AI approach.
grids-in-europe-and-apac.html. vehicle charging and discharging J. Energy Storage 26, 101001.
coordination on distribution network using
169. Bose, B.K. (2017). Artificial intelligence multi-objective particle swarm optimization 192. Lee, Z.J., Li, T., and Low, S.H. (2019). ACN-
techniques in smart grid and renewable and fuzzy decision making. Energies 9, 1–17. Data: Analysis and Applications of an Open
energy systems - some example applications.
EV Charging Dataset, In Proceedings of the
Proc. IEEE 105, 2262–2273. 181. Bai, X., and Qiao, W. (2015). Robust
Tenth International Conference on Future
optimization for bidirectional dispatch
170. Frost&Sullivan. (2020). Innovations in smart Energy Systems, Phoenix, Arizona. https://ev.
coordination of large-scale V2G. IEEE Trans.
metering enabling modernization. https:// caltech.edu/dataset.
Smart Grid 6, 1944–1954.
store.frost.com/innovations-in-smart-
metering-enabling-grid-modernization.html. 182. Su, S., Li, H., and Gao, D. (2017). Optimal 193. Zhang, D., Han, X., and Deng, C. (2018).
planning of charging for plug-in electric Review on the research and practice of deep
171. BIS Research (2019). Global artificial vehicles focusing on users’ benefits. Energies learning and reinforcement learning in smart
intelligence in energy market: focus on 10, 952. grids. CSEE J. Power Energy Syst. 4, 362–370.
product type, industry, applications, funding -
analysis and forecast, 2019–2024. https:// 183. Yang, Q., Li, J., Cao, W., Li, S., Lin, J., Huo, D., 194. Barton, B., and Schütte, P. (2017). Electric
www.researchandmarkets.com/reports/ and He, H. (2020). An improved vehicle to the vehicle law and policy: a comparative analysis.
4900299/global-artificial-intelligence-in- grid method with battery longevity J. Energy Nat. Resour. Law 35, 147–170.
energy-market. management in a microgrid application.
Energy 198, 117374. 195. Madina, C., Zamora, I., and Zabala, E. (2016).
172. Sullivan, F. (2016). Future of the Smart Grid. Methodology for assessing electric vehicle
The Global Smart Grid Industry will Reach 184. Shipman, R., Naylor, S., Pinchin, J., Gough, R., charging infrastructure business models.
$112.7 Billion in 2025 growing at 6.2% (Frost & and Gillott, M. (2019). Learning capacity: Energy Policy 89, 284–293.
Sullivan). https://store.frost.com/future-of- predicting user decisions for vehicle-to-grid
the-smart-grid-industry.html. services. Energy Inform 2, 1–22. 196. Wood, E.W., Rames, C.L., Bedir, A.,
Crisostomo, N., and Allen, J. (2018). California
173. Wadhera, A., Ayoub, J., and Roy, M. (2018). 185. Soares, J., Sousa, T., Morais, H., Vale, Z., and
plug-in electric vehicle infrastructure
Smart Grid in Canada (Natural Resources Faria, P. (2011). An optimal scheduling
projections: 2017-2025 - future infrastructure
Canada). https://www.nrcan.gc.ca/sites/www. problem in distribution networks considering.
needs for reaching the state’s zero emission-
nrcan.gc.ca/files/canmetenergy/pdf/Smart% In IEEE SSCI 2011 - Symposium Series on
vehicle deployment goals. https://doi.org/10.
20Grid%20in%20Canada%20Report%20Web Computational Intelligence - CIASG 2011:
2172/1430826.
%20FINAL%20EN.pdf. 2011 IEEE Symposium on Computational
Intelligence Applications in Smart Grid,
174. Department of Energy Grid Modernization 197. Huang, Y., and Kockelman, K.M. (2020).
pp. 25–32.
and the Smart Grid (2020). https://www. Electric vehicle charging station locations:
energy.gov/oe/activities/technology- 186. Ghofrani, M., Arabali, A., Etezadi-Amoli, M., elastic demand, station congestion, and
development/grid-modernization-and-smart- and Fadali, M.S. (2014). Smart scheduling and network equilibrium. Transp. Res. D 78,
grid. cost-benefit analysis of grid-enabled electric 102179.
vehicles for wind power integration. IEEE
175. Tan, K.M., Ramachandaramurthy, V.K., and Trans. Smart Grid 5, 2306–2313. 198. Valogianni, K., Ketter, W., Collins, J., and
Yong, J.Y. (2016). Integration of electric Zhdanov, D. (2020). Sustainable electric
vehicles in smart grid: a review on vehicle to 187. Li, W., Rentemeister, M., Badeda, J., Jöst, D., vehicle charging using adaptive pricing. Prod.
grid technologies and optimization Schulte, D., and Sauer, D.U. (2020). Digital Oper. Manag. 29, 1550–1572.

2322 Joule 5, 2296–2322, September 15, 2021